Physics-guided machine learning framework for performance and economic assessment of a pico-hydroelectric system

Pico-hydropower offers an effective solution for high-head, low-flow rural electrification; however, most existing
studies examine hydraulic performance, machine-learning prediction, and techno-economic feasibility in isolation.
This study develops an integrated physics-guided machine learning framework that combines laboratory
turbine characterization, physically constrained regression modeling, and design-oriented techno-economic
assessment for a 3 kW Pelton-type pico-hydroelectric system. Experiments conducted under heads of 40–80 m
and discharges of 1.2–5.6 L s⁻¹ generated a 240-sample dataset used to train and validate five predictive models:
a physics-based baseline, Ridge Regression (RR), Random Forest, Gradient Boosting, and a tuned Multi-Layer
Perceptron. Using five-fold cross-validation and a 20 percent independent test set, RR achieved the highest accuracy
(RMSE = 0.045 kW, R² = 0.992), with bootstrap uncertainty analysis confirming narrow confidence
intervals and strong robustness under limited data. Coupling the validated RR model with hydraulic constraints
enabled extrapolation to the 100 m design head, yielding a predicted output of 2.93 kW at 5 L s⁻¹, closely
matching physics-based estimates. The techno-economic assessment produced an LCOE of 0.036 USD kWh⁻¹ and
demonstrated positive NPV across sensitivity ranges. The proposed framework provides a physically interpretable,
statistically reliable, and computationally efficient tool for performance prediction and feasibility evaluation
of small-scale hydropower systems.